aboutsummaryrefslogtreecommitdiffstats
path: root/docs/design
Commit message (Collapse)AuthorAgeFilesLines
* Shape Khmer and Myanmar as HarfBuzz doeseunhhu2026-09-281-1/+12
| | | | | | | | | | | | Two more shapers on the syllabic groundwork, from HarfBuzz 10.2's source. Khmer reorders before any lookup: a coeng and Ro go before the base, masked for pref, what follows them for cfar, a left matra goes first; split matras decompose with U+17C1. Myanmar finds the base after a kinzi, gives each glyph its place around it and sorts by place, and zeroes mark advances before GPOS rather than after. Noto Sans Khmer and Noto Sans Myanmar shape as HarfBuzz does at 20,000 runs a plan, and so do Unifont's Khmer and Myanmar (no GSUB).
* Shape the nine Indic scripts as HarfBuzz doeseunhhu2026-09-281-1/+38
| | | | | | | | | | | | | | | | | | | | | | | | | | The Indic shaper, from HarfBuzz 10.2's source: syllables found by the Indic machine; each syllable's base consonant, its pre-base matras and reph moved around it and masked for the basic features, applied one at a time within the syllable; then matras, reph and pre-base-reordering consonants placed again by what the features made. Vowel constraints, Indic decompositions and compositions, dotted circles for broken syllables, and per-syllable matching come with it. Plans carry GSUB pauses (steps where the shaper does its part) and the stages of the features the shaper asks whether they would substitute; the runtime answers with each subtable's would_apply, as HarfBuzz does. Masks are now the common ones and each shaper's own. text/syllabic.py reads HarfBuzz's Ragel grammars (Indic, Khmer, Myanmar) and compiles each to a minimal DFA with Ragel's scanner semantics; beside the machines Ragel generated from the same grammars, 100,000 random runs each find no syllable that differs. It also reads HarfBuzz's Indic table (Unicode 16) cell for cell. Every installed font shapes as HarfBuzz does (213 fonts, 1,688 plans); the Indic ones at 20,000 runs a plan (FreeSerif's eight scripts, FreeSans, Unifont's nine without GSUB), and nine new-spec Noto Sans fonts at 5,000.
* Shape Hangul, kern with kern tables, and place marks without GPOS, as ↵eunhhu2026-09-281-12/+25
| | | | | | | | | | | | | | | | | | | | | | | | HarfBuzz does The planner now picks who positions as HarfBuzz 10.2 does: GPOS, the old kern table when GPOS has no kern feature, and marks placed by combining class and ink boxes when there is no GPOS; the Hangul shaper composes syllables the font has whole, takes apart those it lacks into jamo for ljmo, vjmo and tjmo, and moves tone marks before their syllable. The section grows to an 80-byte header (the kern table, and each glyph's ink box when a plan places marks without GPOS) and 88-byte plans (the kern mask, the jamo masks). Kern tables are applied pair subtable by pair subtable with HarfBuzz's own binary search; state machines, class subtables and cross-stream kerning, which no installed font has, are refused. The check spells Hangul syllables every way text does. Every installed font Vega accepts shapes as HarfBuzz does (209 fonts, 1,520 plans, a thousand runs each), and the fonts refused before at 20,000 runs a plan: Noto Sans CJK KR and Baekmuk for Hangul, IPA Gothic and Noto Mono for marks without GPOS, FreeSans, FreeSerif, Liberation and DejaVu Sans ExtraLight for kern tables, Unifont's Latin, Greek, Cyrillic, Hebrew and Hangul.
* Speak AT-SPI2 on the accessibility bus from the display, over Vega's own D-Buseunhhu2026-09-281-0/+32
| | | | | | | | | | | | | | | | vega display --access joins the accessibility bus as AT-SPI2 objects for screen readers such as Orca, with no toolkit, ATK or libdbus: dbus.c speaks the D-Bus wire format (EXTERNAL authentication, headers, aligned bodies, calls, replies, signals). The tree follows what the say surface says: a heading, a label per field, a list box per list whose items know their place and which is selected and focused, tables, the entry, the keys as buttons; answers come from the app when asked, one object has the focus, and focus moves are said as events. Assistive technology can act too: Selection focuses a row, Action presses a key. The live loop now draws only when something changed, so a screen reader's questions no longer cost frames. The check walks the tree with gdbus on a private session bus with at-spi-bus-launcher, and follows the focus through keys, Selection, Action and the events.
* Make the web surface accessible the way the say surface is, checked in ↵eunhhu2026-09-281-0/+25
| | | | | | | | | | | | | Chromium's accessibility tree Lists are listboxes that keep the keyboard focus, the focused row their active descendant, each option with its place in the whole list though only the visible rows are drawn; tables are grids with row indices; tones other than normal and muted are said, not only coloured; keys are named buttons. The page now updates in place instead of rebuilding each list, so the focus a screen reader follows survives every redraw. The browser test reads Chromium's accessibility tree: roles, names, focus, selection and the active descendant moving within the same list.
* Set paragraphs into lines: bidi, script runs, shaping, line breaking, visual ↵eunhhu2026-09-281-0/+27
| | | | | | | | | | | | | | | order text/layout.c puts the checked parts together: a paragraph's levels, scripts resolved for Common and Inherited characters, break opportunities, one shaping to measure, lines broken at the last opportunity that fits (trailing spaces hang, and a line that shapes wider than measured backs off), then per line L1, runs of one level and one script shaped again, the runs in L2's order and their glyphs placed. test-layout checks 30,000 random mixed-script paragraphs: lines cover the text, mandatory breaks end lines, lines fit when they can, and glyphs come in the order UAX #9 gives the characters. vega-symbols paragraph draws a layout through the symbol file's glyphs.
* Break text into lines, graphemes, words and sentences as Unicode's tests requireeunhhu2026-09-281-0/+19
| | | | | | | | | UAX #14 line breaking (LB1-LB31, numbers tailored as Unicode's tests are) and UAX #29's grapheme cluster, word and sentence boundaries, each from generated property tables, the rules taken in order, in linear time. They pass all of Unicode 15.1's LineBreakTest (10,274 cases), GraphemeBreakTest (1,187), WordBreakTest (1,826) and SentenceBreakTest (512).
* Shape text as HarfBuzz does, from a symbol file, checked against HarfBuzzeunhhu2026-09-281-3/+58
| | | | | | | | | | | | | The build tool compiles a font's shaping plans the way HarfBuzz's planner does (features, stages, masks, lookups per script and direction) and carries them with the font's character map, advances and GDEF, GSUB and GPOS tables in a section of the symbol file (VEGASYM4); glyphs shaping reaches become symbols of their own. A small engine ports HarfBuzz 10's default, Arabic and Hebrew shaping step for step and reads the tables in place, every read bounds-checked. The build refuses a file whose shaping differs from HarfBuzz's on random runs: DejaVu Sans agrees on 600,000 runs in six scripts both ways, and every installed font the tool accepts agrees too, at 1.05-1.18 times HarfBuzz's time.
* Normalize text to NFD and NFC, passing all of Unicode's normalization testseunhhu2026-09-281-0/+9
|
* Resolve bidirectional text as UAX #9 does, passing all of Unicode's bidi testseunhhu2026-09-281-0/+42
|
* Take keyboards, mice and touch screens on the display (evdev)eunhhu2026-09-271-2/+15
| | | | | | | | | | | | | | vega display --live reads the kernel's input events with no library: keys become Pure's key names and the text they type on a US layout, going through the TUI's own handling (the input line takes text, else the app takes keys); a mouse moves an outlined pointer and clicks, a touch screen presses where it is touched, scaled from the device's range; the line pressed is the row. --input FILE reads recorded events to their end and prints the last frame, which is how this screenless Pi checks it: after typing two items with Shift, Enter, Tab, Space and a touch on a row, the display's last frame is the TUI's after the same keys and click, and came back from the display controller as drawn.
* Put the frames on a display through KMS, checked by writebackeunhhu2026-09-271-2/+17
| | | | | | | | | | | | | | | vega display draws the same frames on a display with the kernel's mode setting and no library between: a connected connector's mode on a CRTC, its primary plane flipped between two dumb buffers by atomic commits, the grid as many cells as fit. Frames are drawn in cached memory and copied over, since blending over write-combined buffers read them back slowly (a 1080p frame took 370 ms; now 20.6 ms with a vertical sync and the check below). This Pi has no screen but its display controller has writeback connectors: --writeback sets a mode on one and takes every frame back as the controller put it out. Each came back exactly as drawn, and the new check finds it the pixel surface's frame for the same grid, pixel for pixel.
* Draw the TUI's cells as pixels with the symbols' glyphseunhhu2026-09-272-3/+46
| | | | | | | | | | | | | | | vega pixels runs the same scripts as tui --dump and writes each frame as an image: the view's cells (only those on screen are evaluated) become glyphs, each drawn once per size with the exact area of its curves, in its tone's colour on its row's background. No page, no scene in between. A 1920x1350 grid takes 12 ms a frame here, writing the image included. Symbol files carry the font's line now (format 3: ascent and descent where format 2 kept zeros), the cell height and baseline; and the build tool takes fonts in order, later ones filling the codes earlier ones lack: a CJK font alone drew ambiguous-width arrows two cells wide in a one-cell grid, which the new check caught, so a monospace Latin font comes first and Hangul after.
* Draw text from an atlas the CPU fills, composed by the GPUeunhhu2026-09-271-16/+24
| | | | | | | | | | | | | | | | | Each size's symbols are drawn on the CPU into one 8-bit atlas, every pixel the exact area its curves cover, and text is a quad per glyph reading it. A 1080p page of 11,382 glyphs composes in 1.07 ms at 16 px (it took 525 ms computing coverage per fragment), and the composed pixels are the CPU raster's exactly (worst difference 0 for DejaVu's ASCII, 1 for Noto's 11,172 Hangul syllables). Computing coverage on the GPU lost on both counts: a compute pass filling the atlas took 5-13x longer than one Cortex-A76 core (35 against 2.7 ms at 16 px), and its faster sorting-network form computed whole rows wrongly under this Mesa V3D. Coverage now has one implementation, in C, and the shaders only compose. vega_symbol_raster takes a row stride to draw into the atlas in place.
* Name Vega's parts: lib/, surfaces/ and symbols/eunhhu2026-09-272-3/+3
| | | | | | | src/ held two parts, the surfaces that show modules to people and the symbol format with its CPU and GPU drawing. As in Pure (design 009), top level directories are now the parts, and the README says what each builds.
* Draw symbols on the GPU from the point table as it iseunhhu2026-09-271-1/+26
| | | | | | | | | | | | | | | | OpenGL ES 3.1: the file's points go up as one storage buffer, each symbol gets 16 row bands listing the curves they reach (cut by integer division of doubled coordinates, so CPU and shader agree), and every pixel of a symbol's box sums the exact area of its bands' curves, as vega_coverage does. vega-symbols gpu draws every symbol against the CPU reference and times 1080p pages. On the Pi 5's V3D it matches the reference (mean 0.001-0.1 levels, worst 4) but takes 142-757 ms a page, 83% of it in the area function. A sorting network in registers made it 2.3x faster yet computes wrong rows in the fragment stage with this driver, while the same code is exact in a compute shader; so drawing each glyph size once into an atlas by compute comes next.
* Store symbol outlines as point chains, and keep glyph advances wholeeunhhu2026-09-271-36/+54
| | | | | | | | | | | | | | | | | | | | Format 2 (VEGASYM2) writes each contour as a chain of 4-byte points, on or off the curve as in TrueType: endpoints are shared, lines have no control point, and two off points imply the on point between them. A curve is still had from one point and its two neighbours, so a GPU can point at curves by point index and upload the table as it is. Cubics (CFF, SVG curves and arcs) become quadratic splines the cu2qu way, joins left implied, within a unit. Files are about half of format 1: 1.3-1.4x a TrueType font's glyf (the whole subset font's size), 2.0-2.9x CFF, 2.6x smaller than the SVG icons. Opening checks each symbol's first and last point instead of every curve: 0.34 ms rather than 19 ms for 11,172 Hangul syllables. Points decode without branching on their flags (9.7 ns a point), still 1.2-10x cheaper than FreeType's outline. Glyph advances were half their width since outlines are doubled for FreeType's implied points; they now match hmtx, and the check says so.
* Draw glyphs and icons from one symbol format, by exact pixel areaeunhhu2026-09-271-0/+84
| | | | | | | | | | | | | | | | | | | Symbol files (VEGASYM1) hold glyphs and icons as quadratic curves, verified once per symbol and then read in place. vega-symbols builds them from fonts (through FreeType, losslessly: twice the font's units keeps TrueType's implied points whole) and from solid SVG paths (arcs included), refusing by name what an icon asks for that symbols cannot hold. The reference rasterizer computes each pixel's covered area exactly: every curve's integral of clamp(1 - x) dy inside the pixel, split where it crosses the pixel's sides. The Slug-style two-ray average first tried missed FreeType by 15.4 levels on average at 12 px; exact area is within 1.2-3.5 at 8-16 px, and where it differs most it matches 64x64 supersampling, not FreeType. Reaching an outline costs 4.6-29x less than through FreeType, but the files are 2.1-6.6x larger than glyf or CFF: design 002 has the numbers and the next question (points shared along contours, lines without control points).
* Add a board used by people and programs alikeeunhhu2026-09-261-1/+3
| | | | | | | | | | examples/board.aup is a kanban board: people add and move cards on Vega's surfaces, programs post JSON to the same module served by pure-serve, with no routes written. The checks write cards with the CLI into a store, serve that store over HTTP where a program moves one and adds another, and read it again with the CLI, the focus still on the card it was on.
* Give live surfaces a clock and run timers; add a stopwatcheunhhu2026-09-261-1/+11
| | | | | | | | | | | | | The terminal and the web page set the time events see to the wall clock and tick an app's timers on time: each round ticks what is due, draws, then waits for input, a frame or the next tick, and a tick is news to every open page at once. Scripts have no clock: time MS sets the time the next events see and tick TIMER ticks, so an app that uses time replays exactly. A module swap reschedules the timers. examples/stopwatch.aup runs on every surface; the checks drive it by script, on a pseudo-terminal by the clock, and on the web page.
* Add a tally example: a keyed list on every surfaceeunhhu2026-09-261-1/+1
| | | | | | | | examples/tally.aup counts typed words in a list keyed by the word: the screen stays sorted with each word once, typing a word again counts it and moves the focus to it. It runs on the three surfaces with the other examples, and vega check finds nothing in it.
* Swap a running app's module: --watch and the swap commandeunhhu2026-09-261-0/+24
| | | | | | | | | | | | | vega tui|web MODULE --watch replaces the running app's module when its file is written again (aup watch writes it); states keep their values by name through pure_app_replace and every web page hears of it. The file is compared by the digest of its bytes, at most five times a second, so a module of the same size written within the same second is not missed; a module that does not load leaves the app as it was and is not tried again. Modules alternate between two slots so a failed load never touches the running one. Scripts get swap MODULE; whether a module runs frames is now a property of the screen.
* Restate the bytes a changed frame sendseunhhu2026-09-261-1/+1
| | | | | | The todo example now counts its open tasks, so the field line changes with more events; a changed frame sends 143 B on average, not 139 B.
* Show filtered lists and their shared focuseunhhu2026-09-261-0/+8
| | | | | | | | | | | | | | | | A list may now be filtered (each t in tasks where !t.done), and several filtered lists over one list share one focus: the focus view names an item, so Vega asks Pure which row shows it and highlights it only in the list that holds it; the other lists have no focus (null in the web surface's JSON, and the spoken surface names the item where it is). A click or point on a row points at the item that row shows, and the wheel on a list without the focus focuses the row under it. Pure's rows now take a mutable app. examples/today.aup shows what is open and what is done; its CLI output matches a model over random adds, moves, toggles and points into either list, and the web surface leaves the other list without a focus.
* Add a web surface: the same spec as a pageeunhhu2026-09-261-1/+25
| | | | | | | | | | | | | | | | | | | vega web MODULE serves the screen to a browser (127.0.0.1 unless given --listen): one dependency-free page draws the title, fields, lists as listboxes and tables, with tones as colors and the focus highlighted. It asks for the screen as JSON with the rows it can show per list or table, so only those cells are evaluated, scrolled like the TUI around the focus. Keys go as KeyboardEvent.code, which is Pure's key naming already; typed text commits with Enter; a click or the wheel moves the focus through pure_app_point; an event stream tells every open page of each change. Scrolling, key presses and text commits are now shared helpers. The checks drive it against the model (1,200 random steps: title, field, rows, focus, the window of items with badges and tones), check a table's columns and cells, post at random to every example, and in headless Chromium over the DevTools protocol press keys, click a row and type Korean text into the page itself.
* Move the focus from a surface: clicks, the wheel, pointeunhhu2026-09-261-2/+8
| | | | | | | | | | | | | | | | | Only the app's key handlers could move a list's focus. When the focus view is a plain index position (`view tasks_list_focus = at.position`), Vega now asks Pure for that index and moves it as a user pointing at a row (pure_app_point), with no code in the app. The grid remembers which item each line shows; in the TUI a click on a row focuses it and the wheel moves the focus one row (SGR mouse reports, turned off on exit). Scripts say point VIEW ROW or click LINE COLUMN on the last shown frame. The focus is stored like any event, so the next run of any surface starts from it. The model now points too: random scripts with points (some past the end) and TUI clicks still match it on every surface, a mouse click on a real pseudo-terminal moves the focus, and the stored session keeps it.
* Send the terminal only the cells that changedeunhhu2026-09-261-2/+13
| | | | | | | | | | | | | | | The TUI rebuilt its grid from the app on every change and then sent all of it. It now keeps what the terminal shows and sends each run of changed cells after a cursor move (short unchanged gaps stay in a run, being cheaper than another move; a run that starts in the second half of a wide character starts at its first). A resize sends everything again. vega tui --paint WxH writes the bytes of each frame; the checks replay them with a small terminal emulator and compare with the dumped frame: 4,800 random frames at three sizes agree, and a change costs 139 bytes on average against 1,220 for a whole frame. The interactive check now reads the final screen from the replayed bytes.
* Add vega check for conventions a spec misseseunhhu2026-09-261-1/+25
| | | | | | | | | | | | | Surfaces read a view that misses a convention as something else without a word: a number label makes a list a table, a text tone disappears, a list never shows its other columns, X_focus without a table becomes a field, a second text input has nowhere to type. vega check MODULE names each such place (the view or input and what the surface does instead) and exits 1; the rules sit next to the code that reads the screen. The checks build a spec that misses eight conventions and expect exactly its nine findings, and none in the examples.
* Keep a surface's app in a storeeunhhu2026-09-251-0/+8
| | | | | | | | | | | Every surface takes --store PATH and runs the app on Pure's tree store, so a list typed in the terminal is there the next time, on any surface. Vega and the app carry no saving code; Vega saves the last frame changes when it ends. Checks let the CLI and the sentence surface take turns on one store and match the model each time, and reopen an interactive session's store.
* Add the first component and three surfaceseunhhu2026-09-251-0/+66
Vega reads a verified Pure module by convention: a table view with a label column is a list (with optional badge, detail and tone), any other table is a table, <table>_focus is its focused row, a title view is the title and other views are fields. lib/vega.aup gives aup apps the tones, which say what a row means; each surface decides how it looks or sounds. vega tui draws the screen in a terminal, keeps the focused row in view and evaluates only the cells on screen; vega cli prints plain lines for pipes; vega say speaks the screen and each focus change as sentences. Display width handles Hangul, CJK, emoji, combining marks and controls. tests/check.py compares the CLI and NUI output with a reference model of examples/todo.aup line for line, checks 9,600 TUI frames for scrolling, focus, badge alignment and the number of cells evaluated, drives the interactive TUI on a pseudo-terminal, and runs every aup example on every surface.
generated by cgit v1.2.3 (git 2.46.0) at 2026-10-20 15:14:19 +0000